Detecting early disease, e.g., Alzheimer's disease (AD), all cause dementia, frailty, institutionalization, etc., is a crucial step towards developing interventions for preventing or delaying disease, and is an important goal in many aging studies. For longitudinal studies, time- dependent receiver operating characteristics (ROC) analysis, which integrates the time dimension in ROC analysis, is a useful and powerful approach to evaluate the prognostic capacity of markers for predicting incident disease during various length of follow-up. However, an intrinsic assumption in most commonly used models is that the censoring process is random , an assumption violated for censoring due to death, since death is associated with disease. Censoring due to drop out is also often non-random as it is been to associate with poor health and negative outcomes such as accelerated cognitive decline in aging researches. Ignoring the informative censoring for disease may lead to erroneous and misleading results. Furthermore, there is substantial evidence that the variability, in addition to the mean, of the markers is associated with disease. Thus it is important to utilize the heterogeneous variance to improve diagnostic accuracy. However, most ROC analyses ignore it which might result in inaccurate or misleading diagnosis. Statistical methods for handling informative censoring and taking advantage of heterogeneous variance in markers in time dependent ROC analysis is lacking. We propose a time-dependent ROC approach that makes use of the heterogeneous variance in markers and takes non-random censoring into account to identify early disease. The effect of aging related genetic polymorphisms (e.g., ApoE4, CETPV405V), and social and behavioral covariates (i.e., perceived stress, personality, depression, and anxiety) will also be examined. We anticipate that our proposed methods, which will be applied to data from the EAS, will guide future efforts to prevent aging related disease by providing more accurate identification of disease onset so that prevention or treatment procedures can be applied more effectively. In addition, these approaches will have broad applicability given that all longitudinal aging studies share similar heterogeneous variance and informative censoring issues.